AI LoRA Trainer API — Train Custom LoRA Models
Live8 trainersMuapi's AI LoRA trainer API lets you fine-tune a custom LoRA (Low-Rank Adaptation) model from your own image dataset — no GPU rental, no training infrastructure to manage. Upload a zip of images, optionally with per-image text captions, pick a trigger word and step count, and get back a downloadable .safetensors file ready to plug into generation. Eight dedicated training endpoints cover five base model families — Flux.1 [dev], FLUX.2 Klein 4B/9B, Qwen-Image, Qwen-Image-2512, and Z-Image / Z-Image Base — all via the same REST pattern, pay-per-step or flat-fee, from $2.00 per 1,000 steps.
Flux LoRA Trainer
Train a custom Flux LoRA from a zip of 10-50 images and a trigger phrase. Flat fee per run regardless of dataset size — reuse the resulting .safetensors across unlimited Flux Dev generations.
FLUX.1 [dev] Style LoRA Trainer
Style-focused LoRA training on FLUX.1 [dev]. Zipped dataset in, downloadable safetensors output, instant platform AIR deployment for downstream inference.
FLUX.2 [klein] 4B Style LoRA Trainer
Style-focused LoRA training on the smaller, faster FLUX.2 Klein 4B architecture. Zipped dataset in, downloadable safetensors output, instant AIR deployment.
FLUX.2 [klein] 9B Style LoRA Trainer
Style-focused LoRA training on the larger, higher-quality FLUX.2 Klein 9B architecture. Zipped dataset in, downloadable safetensors output, instant AIR deployment.
Qwen-Image LoRA Trainer
Fine-tune Qwen-Image on a zipped dataset for character, subject, or artistic-style consistency. Customizable steps (1,000-10,000), learning rate, and LoRA rank.
Qwen-Image-2512 LoRA Trainer
Fine-tune the newer Qwen-Image-2512 checkpoint from a zipped dataset. Same pay-per-step pricing as the base Qwen-Image trainer, with the updated model's improved fidelity.
Z-Image LoRA Trainer
Train custom Z-Image LoRA adapters from a zip archive with auto-tuned hyperparameters. Customizable training steps from 1,000 to 10,000, adjustable learning rate, and flexible LoRA rank.
Z-Image Base LoRA Trainer
Fine-tune the full (non-distilled) Z-Image Base checkpoint from your own dataset. Auto-tuned defaults with full manual control over steps, learning rate, and rank.
What is an AI LoRA Trainer API?
A LoRA (Low-Rank Adaptation) is a small, portable weights file that teaches a base image model a specific character, subject, brand identity, or visual style — without retraining the full model. An AI LoRA training API turns that process into a single HTTP call: upload a zip of reference images, submit the request, and poll for a downloadable .safetensors file once training completes.
Muapi exposes eight LoRA trainer endpoints through the same unified REST pattern used across every model on the platform — one API key, one request/poll flow, pay-as-you-go pricing. Train once, then reuse the resulting LoRA across unlimited generations on the matching base model, or download it for use in your own pipeline.
Key Capabilities
Zip Dataset Upload
Bundle 4-50 training images (plus optional .txt caption files) into a single zip and upload it to a publicly reachable URL — no dataset infrastructure to manage.
Custom Trigger Words
Pick a unique trigger phrase that activates your trained concept in any future prompt, without colliding with the base model's existing vocabulary.
Five Base Model Families
Train on FLUX.1 [dev], FLUX.2 Klein 4B/9B, Qwen-Image, Qwen-Image-2512, or Z-Image / Z-Image Base — pick the checkpoint that matches your downstream generation pipeline.
Adjustable Steps, Rank & Learning Rate
Auto-tuned defaults get you a usable LoRA fast, with full manual control over training steps (1,000-10,000), learning rate, and LoRA rank for advanced tuning.
Downloadable Safetensors Output
Every trainer returns a standard .safetensors file you can download and reuse in any compatible generation pipeline — not locked to Muapi.
Instant Platform Deployment
Trained LoRAs register instantly for downstream inference on Muapi — no separate publish step before you can start generating.
LoRA Trainer Comparison
| Trainer | Base Model | Price | Best For |
|---|---|---|---|
| Flux LoRA Trainer | FLUX.1 [dev] | $3.20 / run | Subjects & styles, flat per-run pricing |
| FLUX.1 [dev] Style LoRA Trainer | FLUX.1 [dev] | $2.50 / 1,000 steps | Style-focused training with per-step billing |
| FLUX.2 [klein] 4B Style LoRA Trainer | FLUX.2 Klein 4B | $4.50 / 1,000 steps | Faster iteration on the smaller Klein architecture |
| FLUX.2 [klein] 9B Style LoRA Trainer | FLUX.2 Klein 9B | $4.00 / 1,000 steps | Higher-quality style LoRAs on the larger Klein model |
| Qwen-Image LoRA Trainer | Qwen-Image | $2.00 / 1,000 steps | Lowest-cost per-step training |
| Qwen-Image-2512 LoRA Trainer | Qwen-Image-2512 | $2.00 / 1,000 steps | Same low cost on the newer Qwen checkpoint |
| Z-Image LoRA Trainer | Z-Image | $2.50 / 1,000 steps | Fast iteration with auto-tuned defaults |
| Z-Image Base LoRA Trainer | Z-Image Base | $2.50 / 1,000 steps | Full (non-distilled) checkpoint fine-tuning |
How to Train a Custom LoRA via API
- Prepare your dataset — collect 4-50 high-quality images of your subject or style. Vary angles, expressions, and lighting for a subject; keep the aesthetic consistent for a style.
- Optionally add captions — include a
.txtfile per image with the same base name, using a placeholder for your trigger word. - Zip and host the archive — bundle everything into a single zip at a publicly reachable URL, or upload via
POST /api/v1/upload_file. - Submit the training request —
POST /api/v1/{trainer-slug}with your zip URL, trigger word, and optional steps/learning-rate/rank overrides. - Poll for completion — check
GET /api/v1/predictions/{request_id}/resultuntilstatusiscompleted, then download the.safetensorsoutput. - Generate with your trigger word — pass the trained LoRA and your trigger word into the matching base model's generation endpoint to reproduce your subject or style.
Frequently Asked Questions
What is the AI LoRA Trainer API?
Muapi's AI LoRA trainer API lets you train a custom LoRA (Low-Rank Adaptation) model from your own image dataset via a single REST endpoint. Upload a zip of images, optionally with text captions, pick a trigger word, and receive a downloadable .safetensors LoRA file you can reuse across unlimited generations.
Which base models can I train a LoRA on?
Flux.1 [dev], FLUX.2 Klein 4B and 9B, Qwen-Image, Qwen-Image-2512, and Z-Image / Z-Image Base — eight dedicated training endpoints across five model families.
How much does LoRA training cost?
From $2.00 per 1,000 training steps on Qwen-Image and Qwen-Image-2512, $2.50 per 1,000 steps on Z-Image / Z-Image Base and FLUX.1 [dev] Style, up to $4.50 per 1,000 steps on FLUX.2 Klein 4B. The Flux LoRA Trainer is a flat $3.20 per training run regardless of dataset size.
How many images do I need to train a LoRA?
Most trainers accept 4-50 images per zip archive. 10-30 well-varied images typically give the best results — different angles, lighting, and expressions for a subject; consistent aesthetic for a style.
What do I get back when training completes?
A downloadable .safetensors weights file, plus instant registration on Muapi for downstream inference — no separate publish step required before generating with it.
Can I use the trained LoRA outside of Muapi?
Yes — the .safetensors output is a standard LoRA weights file compatible with any pipeline that supports the matching base model architecture, not locked to Muapi's inference.
Can I get AI LoRA training API access right now?
Yes. Sign up at muapi.ai, create an API key from your dashboard, and start training immediately — no waitlist required.